Foam-Cup Stacking
Teleoperated LEAP Hand + Franka stacks foam cups while avoiding deformation and grasp failure.
CoRL 2026 Submission
Current as Touch
Proprioceptive Contact Feedback for Compliant Dexterous Manipulation
University of North Carolina at Chapel Hill
Abstract
Compliance is essential for dexterous manipulation, yet tactile and force sensors can be costly, fragile, and difficult to integrate on low-cost robot hands. Current as Touch learns tactile-like contact feedback from motor current and joint states, then predicts a compliance reference position for a standard PD controller. The robot keeps the familiar position-control interface while adapting interaction forces from intrinsic proprioceptive feedback.
Core Idea
Let the hand feel contact through the currents it already measures.
Motor current is closely related to actuator torque, so contact force, object resistance, and load changes leave traces in the hand's own proprioceptive stream. Instead of estimating external wrenches or commanding torque, the model predicts a contact-aware target position. The resulting PD position error produces compliant grasping force without external tactile sensors.
01
Intrinsic sensing
Uses raw motor current and joint states as tactile-like feedback.
02
Position compatible
Outputs compliance reference positions for standard low-level PD control.
03
Policy ready
Works in both human teleoperation and downstream policy learning.
Why Current Works
Current and joint state track contact force across hands.
Measurements on Unitree Dex3 and LEAP Hand show that motor current and joint states vary consistently with externally measured contact force. This motivates learning from current directly as a contact signal rather than adding fragile external tactile hardware.
Real-Robot Tasks
Four contact-rich settings where force mistakes are visible.
The same current-conditioned interface is evaluated on fragile object handling, dynamic load adaptation, sustained surface contact, and thin-object retrieval.
Whiteboard Wiping
Dex3 on Unitree G1 maintains sustained contact while pressing an eraser against an inclined board.
Single-Card Picking
Dex3 + G1 draws exactly one card from a deck, where excessive normal force pulls multiple cards.
Dynamic Bottle Holding
LEAP Hand + Franka holds a bottle as water is poured in, adapting grip as the load changes.
Results
Current feedback improves safety, speed, and policy robustness.
0.0%
Foam-cup deformation for both novice and skilled users with current-conditioned CRP.
100%
Whiteboard wiping success for both users, with shorter completion times.
100%
Stable dynamic bottle holding at 250 g poured water, compared with 16.7% w/o current.
76.9%
Strict single-card success, up from 55.8%; multi-card failures drop to 0.0%.
Demo
Supplementary video
The video contains the task demonstrations and current/no-current comparisons included with the submission.
Place a compressed demo at assets/videos/current-as-touch-demo-web.mp4, or replace this source with a GitHub Release / CDN video URL.
Materials
Paper, supplementary material, and citation.
@inproceedings{ma2026currentastouch,
title = {Current as Touch: Proprioceptive Contact Feedback for Compliant Dexterous Manipulation},
author = {Ma, Chenyang and Yao, Yunchao and Wei, Zhenyu and Li, Ruogu and Szafir, Daniel and Ding, Mingyu},
booktitle = {Conference on Robot Learning},
year = {2026}
}